| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Technical Developments |
1 From the Department of Diagnostic Radiology, Warren Grant Magnuson Clinical Center, National Institutes of Health, Bldg 10, Rm 1C660, 10 Center Dr MSC 1182, Bethesda, MD 20892-1182 (R.M.S., L.M.P., J.D.M.), and the Department of Radiology, Stanford University Medical Center, Stanford, Calif (C.F.B., R.B.J., D.I.G., S.N.). Received August 5, 1999; revision requested September 24; revision received November 8; accepted November 16. Supported in part by the intramural research programs of the Diagnostic Radiology Department, Clinical Center, Bethesda, Md; National Institutes of Health grants 1R01 CA72023 and LM 07033; Silicon Graphics, Mountain View; the Packard Foundation, Los Altos, Calif; the Lucas Foundation, Menlo Park, Calif; and the Phil N. Allen Trust, Menlo Park, Calif. C.F.B. supported in part by an RSNA Research and Education Foundation Scholar Award. Address correspondence to R.M.S. (e-mail: rms@nih.gov).
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
Index terms: Colon, CT, 75.12111, 75.12115, 75.12117 • Colon, neoplasms, 75.3119 • Computed tomography (CT), image display and recording, 75.12115, 75.12117 • Computed tomography (CT), three-dimensional, 75.12117 • Computers, simulation • Phantoms
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
The appropriate way to perform and interpret CT colonographic studies is still in evolution. Findings in a preliminary clinical study suggest that optimal interpretation consists primarily of using the two-dimensional images supplemented where needed with analysis of three-dimensional (virtual colonoscopic) reconstructions (6). Other work suggests that three-dimensional views may increase detection when used either alone (7) or in combination with two-dimensional images (8). Because a typical CT colonographic study consists of many CT scans (300–600 for supine and prone studies depending on technique), it is time-consuming to interpret (9). There is also a need to improve the sensitivity of CT colonography, which in preliminary reports is 75%–83% for polyps 8–10 mm in diameter or larger (6,10). We hypothesized that computer-assisted polyp detection could potentially improve efficiency of interpretation and increase sensitivity. For these reasons, we developed a computer-assisted detection algorithm and tested it in an established phantom model for colonic polyps.
|
Materials and Methods |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
Image Processing
Simulated polyps.—Polyp synthesis and methods for merging the synthetic lesions with the patient's CT data have been described in detail elsewhere (7,11). Briefly, lesions were simulated as spheric structures 10 mm in diameter by using CT simulator software that models the attenuation, pixel dimensions, and partial volume characteristics of the helical CT scanner. Lesions were composited with the subject's CT data by using nonlinear methods that minimize the appearance of artifacts at the interface of the added lesion and the native colon wall and match image noise characteristics between the synthetic lesion and the background data. Ten identical synthetic lesions were inserted in the colon as follows: Random points along a previously computed central path extending from the cecum to the rectum (12) were identified, then, random radial rays were generated from the path points to the colon wall. The x, y, and z coordinates of the intersection of these rays with the colon wall represented the insertion locations for the center of the synthetic lesions. Although there is the possibility that portions of the colonic wall could be hidden from view by folds (and therefore polyps could not be placed into these hidden spots), our preliminary work indicates that approximately 99% of the colonic wall is accessible with this technique. With use of these coordinates, we merged the lesion data with the patient data, aiming for hemispheric final lesions with one-half of their diameter (5 mm) protruding from the colon wall.
Although the polyps were nominally 10 mm in diameter, the effective size of any polyp depends on the proximity of adjacent structures (such as haustra) or technical factors (such as z-axis broadening). In contrast to the 5-mm protrusion used to position the simulated polyps, the effective size is what a colonoscopist would measure (ie, the effective diameter of the polyp). We measured the effective size by picking points on the edges of the polyp and measuring the distance between the points. Points are picked by being selected with a computer mouse to identify the polyp edge on the three-dimensional surface model. The software locates the vertex projecting nearest to the selected point and uses that information to compute the distance.
Polyp detection.—We transferred the CT images of the colons (one with and the other without synthetic polyps) to a computer workstation (Indigo2 workstation with MAXIMUM IMPACT graphics; Silicon Graphics, Mountain View, Calif). We produced a three-dimensional surface-rendered image of the colons by using our research endoscopic software package. This software provides a realistic endoscopic display of the colonic lumen (13–15). The software was originally developed for virtual bronchoscopy and was adapted for use in CT colonography.
Two experiments were performed. In the first experiment, the CT data were converted from 12 to 8 bits with use of a window level of -475 HU and window width of 1,050 HU. The threshold for generation of the surface was -800 HU. In the second experiment, we converted the CT data from 12 to 8 bits with use of a window level of -225 HU and window width of 550 HU. For the second experiment, the threshold for generation of the surface was -300 HU. The parameters used for the second experiment were based on our observations (of transverse CT images and perspective endoluminal reconstruction images of the colon phantom) that they highlighted the profile of the polyps and improved separation of polyps from adjacent folds. The window width and level settings for the second experiment were a modified version of those used in reference 7 for detection of colonic polyps on transverse CT scans.
In both experiments, the window width was set so that data up to 50 HU were preserved. We identified voxels within the colonic lumen with use of region growing and an upper threshold of -180 HU. The lumen was dilated by two voxels to provide data within the colonic wall. To preserve maximum surface detail, triangle reduction (decimation) was not used. Isolated triangles (those not attached to the colon wall) were removed by means of commercially available software (IMEDIT, version 3.0; Innovmetric Software, Sainte-Foy, Quebec, Canada). This procedure took approximately 2 minutes. Surface-rendered images of the colon and polyps were smoothed for purposes of presentation (16).
We divided the CT data set into two overlapping parts (upper abdomen or lower abdomen and pelvis, 4.5 cm overlap) because of its large size (355 images, 178 Mbytes). We analyzed the two parts individually and combined the results. If a polyp was detected by the algorithm in one of the two parts but not in the other in the area of overlap, it was considered a true-positive detection. This situation occurred when a polyp was near or at the edge of the data set in one of the two parts.
Polyp detection was performed by using software with a prototypic automated polyp detector that identifies regions of the colon wall with abnormal shape. This polyp detector is a modified version of a lesion detector previously shown to identify endobronchial lesions successfully (17,18). As in reference 17, the primary shape criterion for the polyp detector is elliptic curvature of the peak subtype. In simpler terminology, this criterion describes areas that protrude inward from the wall of the colon and are circumferentially round (ie, polypoid). The principle behind the method is shown in Figure 1. The faster convolution-based curvature method described in reference 18 was used with a 5 x 5 x 7-mm kernel. The size of the kernel is that used successfully in reference 18 corrected for the voxel dimensions of the current CT data set. Up to this point, the detection of colonic polyps recaps that of detection of polypoid airway lesions (17,18).
|
|
max and min (expressed per centimeter); the dimensionless sphericity ratio S = |(min - max)/H|; and minimum polyp size (expressed in centimeters). The mean curvature corresponds inversely to the size of the polyp. The sphericity expresses the uniformity of the shape of the potential polyp. By setting an upper limit on the sphericity, those portions of the colon wall that are shaped like ridges (curved in one direction and less curved in the perpendicular direction, such as portions of some haustral folds) can be eliminated. Perfect spheres have a sphericity of 0. Sphericity was computed with 
max, min, and 
, which are averages of the corresponding curvature values taken over all vertices in the lesion. Setting the minimum acceptable polyp size reduces the effect of noise. It takes about 30 minutes of processing time to compute the curvatures and apply the shape criteria. The data were processed by one of the authors (R.M.S.), unblinded to the actual polyp location. Sensitivity was determined by counting the number of detected lesions. The number of false-positive polyp detections was determined. Histographic analysis of curvature was performed in an effort to better understand the shape characteristics of the colon.
|
Results |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
|
|
|
|
|
|
|
|
|
|
|
|
10 mm) was 100% (six of six).
|
Of the three false-positive polyps, one was located at the edge of the divided data set in the area of overlap and was readily discarded as an artifact. The other two false-positive cases were plausible polyps that required further analysis by the radiologist. On the basis of our best ability to reconcile results at CT colonography with those at fiberoptic endoscopy, one of the false-positive cases may in fact be the diminutive polyp that was originally found at sigmoidoscopy. It was measured as 4 mm in diameter at colonoscopy and 5 mm in diameter on the transverse CT scan.
We found that all three false-positive cases could be distinguished from the true-positive cases by means of the average Gaussian curvature: K = max x min, where 
is the mean of K of the detected vertices. All polyps and no false-positive cases had average Gaussian curvature, , of 2.2 cm-2 or more. This works because for a sphere with diameter of 1 cm, K = = 4 cm-2.
The second experiment was designed to determine whether the sensitivity for polyp detection would increase if the thresholds were adjusted to enhance the profile of polyps and diminish the profile of haustral folds. With use of the same data set and the second set of thresholds, the sensitivity increased as eight of 10 polyps were detected (all but polyps 7 and 10). With Gaussian curvature, , of 3.3 cm-2 or more and sphericity ratio, S, of 0.8 or less, all eight polyps were identified and all false-positive cases were excluded. The cutoff values for and S are correct because the polyps become slightly smaller (higher curvature) and rounder (lower sphericity) when the isosurface threshold is increased. The lipoma was not detected in the second experiment, in which the higher thresholds were used, probably because it shrunk more than the polyps did.
|
Discussion |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
As currently practiced, analysis at CT colonography is time-consuming. Interpretation times of 15 minutes to 1 hour per patient have been reported (3,9). Computer-assisted diagnostic methods such as ours may help improve efficiency by directing the physician's attention to sites likely to harbor polyps (15). This efficiency involves not only identifying with high sensitivity any suspicious lesions but also minimizing the number of false-positive observations.
The sensitivity of the method used in this study (100%) was comparable to sensitivities reported for colonoscopy (94%), barium enema examination (65%–75%), and CT colonography (75%–83%) for detection of large polyps (6,10,19,20). The sensitivity of our method for all polyps (60%) is lower in part because the effective size of the four missed polyps was only 8–9 mm. We attempted to insert polyps one-half their diameter so they would protrude 5 mm and have an effective size of 10 mm, but visual and quantitative analysis showed the four missed polyps to be less well placed and, thus, effectively smaller. The results were also affected by depth of insertion of the polyps tested. Thus, as a first step, our approach appears promising and worthy of additional study.
We found that we could increase the sensitivity to 80% (eight of 10 polyps) by using thresholds that enhance the surface profile of polyps. These thresholds worked in part by diminishing the surface profile of adjacent haustral folds. It is unknown whether this tactic would reduce the sensitivity of detection of polyps arising entirely from a haustral fold. In addition, the result of thresholding will be affected by data acquisition parameters (collimation, helical pitch, and reconstruction interval) that influence partial volume averaging, and systematic studies of these effects are necessary (21).
The number of false-positive polyp detections (n = 3) is reasonable when one takes into account that there are 355 images in the CT colonographic data set. We found that an additional parameter (Gaussian curvature) could be used to distinguish the polyps from the false-positive cases. When these methods are applied to real polyps, additional approaches to reduce false-positive detections may be necessary. For example, more complex shape criteria derived from curvature or statistical analyses of the curvatures of the vertices that compose the potential polyp (18) could be constructed. Analysis of the curvature histogram may be useful for developing these solutions. Similar approaches might increase the sensitivity.
Measurement of curvature is a standard image processing method in both two and three dimensions. A typical use is analysis of the boundary of a segmented object (22). The application of curvature analyses to virtual endoscopy is relatively recent. In one study, curvature analysis was used to locate aneurysms and stenoses in the aorta (23). In another study, curvature analysis resulted in successful identification of polypoid endobronchial lesions at virtual bronchoscopy with sensitivities of 47%–88% and specificities of 58%–89% depending on the value of an adjustable parameter (the mean curvature threshold) (17). In the latter study, the sensitivity increased by 20%–34% when only larger lesions (>5 mm) were considered. An abstract describing the concomitant use of curvature and wall thickness to identify colonic polyps has also appeared (24). False-positive detections were a problem in that study too, although the number of such false-positive cases was not reported.
When compared with results in the tracheobronchial tree, the sensitivity of this method applied to the colon was lower (17,18). This is in part because the shape of the colonic wall is more complex than that of the airway. For example, the airway is mainly a smooth bifurcating tube with some rippling due to cartilaginous rings. External structures such as the esophagus or vessels can modify its cylindric shape, but these changes are usually in predictable locations. In contrast, the size and shape of the colon varies greatly depending on how well it is distended. Haustral folds are also more variable in appearance than are cartilaginous rings on endoluminal rendered CT images of the tracheobronchial tree.
Our method requires segmentation of the colonic lumen. The segmentation could be performed by a trained technologist. A radiologist can check the adequacy of the segmentation by inspecting a single image, an anteroposterior projection of the surface reconstruction of the colon. The radiologist can use the same image to determine the quality of bowel distention.
This study has several limitations. We studied only one colon. Additional colons need to be studied since the size and distention of the colon affect its curvature. For example, a less distended colon has greater curvature than does a well-distended colon, and its curvature may overlap in an entirely different way with that of the polyps. It may be possible to address this problem by tailoring the shape criteria to the local amount of colonic distention. Collapsed segments of bowel are uninterpretable with this technique, although CT colonographic sensitivity in general is low in collapsed segments. Additional colons need to be studied so that specificity can be computed.
Only a supine CT colonographic study was used. Many researchers advocate the use of both prone and supine CT colonography to improve the yield in collapsed colonic segments.
Only one combination of scanning parameters was used (eg, helical pitch, reconstruction interval or section overlap, collimation). This was not the focus of the current study, but our colon phantom could be used to evaluate the effect of changes in these parameters on polyp detection (25,26). Especially important is evaluation of artifacts that cause rippling on the surface and could affect detection.
The colon phantom has limitations. Although the colon is from a real patient, the polyps are synthetic, and the relationship of their appearance to that of real polyps is unknown. Our observation, however, is that the synthetic polyps are very similar in appearance to many real polyps and experienced radiologists are unable to distinguish them from real polyps on either transverse CT scans or perspective endoluminal reconstruction images (11). All simulated polyps were nominally 10 mm in size and hemispheric. A wider variety of sizes and shapes of polyps must be studied. The effect of location of polyps relative to haustral folds may be important and needs to be studied. These techniques should be tested on real polyps.
In conclusion, we developed a computer algorithm that detected 100% (six of six) of polyps 10 mm or more in size in a colon phantom. Two of the smaller polyps could be detected by applying methods that enhance the edge profile of polyps. Our colon phantom provided an effective laboratory for the study of computer-assisted diagnostic methods for CT colonography. Additional studies need to be performed to determine the optimum shape criteria for detecting real polyps and to minimize the number of false-positive polyp detections.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMaterials and MethodsResultsDiscussionREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  D. Hock, R. Ouhadi, R. Materne, A.-S. Aouchria, I. Mancini, T. Broussaud, P. Magotteaux, and A. Nchimi Virtual Dissection CT Colonography: Evaluation of Learning Curves and Reading Times with and without Computer-aided Detection Radiology, September 1, 2008; 248(3): 860 - 868. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. H. Kim, J. M. Lee, C.-I. Shin, H. C. Kim, J.-G. Lee, J. H. Kim, J. Y. Choi, H. W. Eun, J. K. Han, J. Y. Lee, et al. Effects of Spatial Resolution and Tube Current on Computer-aided Detection of Polyps on CT Colonographic Images: Phantom Study Radiology, August 1, 2008; 248(2): 492 - 503. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
N. Petrick, M. Haider, R. M. Summers, S. C. Yeshwant, L. Brown, E. M. Iuliano, A. Louie, J. R. Choi, and P. J. Pickhardt CT Colonography with Computer-aided Detection as a Second Reader: Observer Performance Study Radiology, December 1, 2007; 246(1): 148 - 156. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 S. H. Kim, J. M. Lee, J.-G. Lee, J. H. Kim, P. A. Lefere, J. K. Han, and B. I. Choi Computer-Aided Detection of Colonic Polyps at CT Colonography Using a Hessian Matrix-Based Algorithm: Preliminary Study Am. J. Roentgenol., July 1, 2007; 189(1): 41 - 51. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. D. O'Connor, R. M. Summers, J. Yao, P. J. Pickhardt, and J. R. Choi CT Colonography with Computer-aided Polyp Detection: Volume and Attenuation Thresholds to Reduce False-Positive Findings Owing to the Ileocecal Valve Radiology, November 1, 2006; 241(2): 426 - 432. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. Shi, P. Schraedley-Desmond, S. Napel, E. W. Olcott, R. B. Jeffrey Jr, J. Yee, M. E. Zalis, D. Margolis, D. S. Paik, A. J. Sherbondy, et al. CT Colonography: Influence of 3D Viewing and Polyp Candidate Features on Interpretation with Computer-aided Detection. Radiology, June 1, 2006; 239(3): 768 - 776. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
C. L. Partain, H.-P. Chan, J. G. Gelovani, M. L. Giger, J. A. Izatt, F. A. Jolesz, K. Kandarpa, K. C. P. Li, M. McNitt-Gray, S. Napel, et al. Biomedical Imaging Research Opportunities Workshop II: Report and Recommendations Radiology, August 1, 2005; 236(2): 389 - 403. [Full Text] [PDF]
|
|
|
|
|
|
A. Ozgun, E. Rollven, L. Blomqvist, S. Bremmer, R. Odh, and A. Fransson Polyp Detection with MDCT: A Phantom-Based Evaluation of the Impact of Dose and Spatial Resolution Am. J. Roentgenol., April 1, 2005; 184(4): 1181 - 1188. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. M. Summers, J. Yao, and C. D. Johnson CT Colonography with Computer-aided Detection: Automated Recognition of Ileocecal Valve to Reduce Number of False-Positive Detections Radiology, October 1, 2004; 233(1): 266 - 272. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
H. Hoppe, C. Quattropani, A. Spreng, J. Mattich, P. Netzer, and H.-P. Dinkel Virtual Colon Dissection with CT Colonography Compared with Axial Interpretation and Conventional Colonoscopy: Preliminary Results Am. J. Roentgenol., May 1, 2004; 182(5): 1151 - 1158. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
A. M. Aisen, L. S. Broderick, H. Winer-Muram, C. E. Brodley, A. C. Kak, C. Pavlopoulou, J. Dy, C.-R. Shyu, and A. Marchiori Automated Storage and Retrieval of Thin-Section CT Images to Assist Diagnosis: System Description and Preliminary Assessment Radiology, July 1, 2003; 228(1): 265 - 270. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
R. M. Summers, A. K. Jerebko, M. Franaszek, J. D. Malley, and C. D. Johnson Colonic Polyps: Complementary Role of Computer-aided Detection in CT Colonography Radiology, November 1, 2002; 225(2): 391 - 399. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 H. Yoshida, J. Nappi, P. MacEneaney, D. T. Rubin, and A. H. Dachman Computer-aided Diagnosis Scheme for Detection of Polyps at CT Colonography RadioGraphics, July 1, 2002; 22(4): 963 - 979. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
J. T. Ferrucci Colon Cancer Screening with Virtual Colonoscopy: Promise, Polyps, Politics Am. J. Roentgenol., November 1, 2001; 177(5): 975 - 988. [Full Text] [PDF]
|
|
|
|
|
|
R. M. Summers, C. D. Johnson, L. M. Pusanik, J. D. Malley, A. M. Youssef, and J. E. Reed Automated Polyp Detection at CT Colonography: Feasibility Assessment in a Human Population Radiology, April 1, 2001; 219(1): 51 - 59. [Abstract] [Full Text]
|
|
|
|
|
|
W. Luboldt, O. Luz, R. Vonthein, M. Heuschmid, M. Seemann, J. Schaefer, D. Stueker, and C. D. Claussen Three-Dimensional Double-Contrast MR Colonography: A Display Method Simulating Double-Contrast Barium Enema Am. J. Roentgenol., April 1, 2001; 176(4): 930 - 932. [Full Text] |
